Wow, there is a lot of misconception in this thread so let me summarize a few points from the dozens and dozens of tests I have done for both horizontal and vertical 3D modeling.
#1 is the 1X to 3X GSD accuracy is in the point cloud, not the orthomosaic. Sharp changes in elevation just don’t show good accuracy in an orthomosaic. In fact I never bother to make the orthomosaic as the point cloud is the truest representation.
#2 a relative accuracy very near to 1X GSD can be achieved without any GCP or scale constraint. Most people don’t believe this to be true but I have the process and repeatable data that proves the capability of Pix4D. I can get sub-mm relative accuracy on top of a cell tower with only using GPS, not even RTK.
#3 all of the above is 100% dependent on how the pictures are taken. In this construction site with at least one building, nadir images just won’t cut it for high accuracy so be sure to use oblique camera angles. I go all the way to 45 degrees but that isn’t always necessary.